SWE-Spot: Building Small Repo-Experts with Repository-Centric Learning
Small Language Models struggle to handle complex codebases because they lack deep, specific knowledge. The proposed Repository-Centric Learning (RCL) shifts training focus from broad tasks to deep mastery of a single repository using a four-unit experience. This results in SWE-SPOT-4B experts that outperform larger models and commercial tools by internalizing the 'physics' of the target code.